Middleware memory engine for LLMs — replaces linear chat history with three parallel structures (Entity Graph, Semantic Tree, Focus Buffer) to retrieve only the most relevant context within a token budget. Works as a Python library or MCP server for Claude Code, Copilot CLI, and other AI tools.